Advanced EEG-Based Fatigue Estimation through Attention-Based Methods in Brain-Computer Interfaces

Abstract

Reliable detection and prediction of mental fatigue using non-invasive electroen- cephalography (EEG) remains a difficulty due to the non-stationary nature of neu- ral signals and high impact of inter-subject variability. Meanwhile deep learning has advanced the field, traditional convolutional and recurrent models struggle to detect long-range temporal dependencies without experiencing high computational costs. This thesis develops and evaluates a comparative framework of advanced sequence modeling architectures, specifically Transformer-based and Selective State Space Models (Mamba), to address these limitations. I propose a hybrid CNN-Attention mechanism to enhance physiological interpretability through atten- tion mapping, alongside a Mamba-based architecture designed for linear computa- tional scaling (𝑂(𝑁 )) suitable for real-time edge deployment. Within the context of EEG fatigue monitoring, performance is determined more heavily by the evaluation protocol than by the selected sequence architecture. Mamba is proven to be a highly viable alternative to the Transformer. A reliable latency ad- vantage is provided by Mamba, and the true speed advantage is likely underestimated due to the simplified implementation developed for the analysis.